The present invention relates generally to a computer system supporting multiple threads, and more specifically, to address expansion and contraction in a multithreading computer system.
As processor speed of computer systems has increased over the past decades, there has not been a proportional increase in the speed in which the memory of such computer systems can be accessed. Thus, the faster the processor's cycle time, the more pronounced is the delay of waiting for data to be fetched from memory. The effects of such delays have been mitigated by various levels of caching, and in recent processors, by multithreading (MT).
MT allows various core resources of a processor to be shared by a plurality of instruction streams known as threads. Core resources can include execution units, caches, translation-lookaside buffers (TLBs), and the like, which may be collectively referred to generally as a core. During latency caused by a cache-miss or other delay in one thread, one or more other threads can utilize the core resources, thus increasing the utilization of the core resources. In a super-scalar processor simultaneous-multithreading (SMT) implementation, multiple threads may be simultaneously serviced by the core resources of one or more cores.
In contemporary hardware platforms, MT is typically implemented in a manner that is transparent to an operating system (OS) that runs on the MT hardware. One aspect of this characteristic is that the OS does not require modification to utilize the MT hardware. However, transparent MT operation with respect to the OS can result in high variability of response time, capacity provisioning, capacity planning, and billing. This variability can occur because the OS is unaware of whether its tasks have exclusive control of a core, or whether its tasks are executing as threads that share a core. By design, the highest capacity for a memory-intensive workload on MT-capable hardware is achievable when there is a high average thread density when the cores are in use. Additional capacity may be due to increased cache exploitation provided by MT. If an OS does not consistently maintain high average thread densities for utilized cores, then the additional overall throughput capacity provided by MT will not be available. For example, if the hardware runs a single MT thread per core when there is low compute utilization and runs with high thread density when there is high compute utilization, then it can be very difficult to determine how much total MT compute capacity is available to the workload. This hardware variability in the MT thread exploitation can lead to variability in both transaction response times and in billing in a similar fashion as previously described with respect to capacity.